Traffic Light Detection Algorithm Based On In-car Camera

The traffic light detection algorithm based on in-car camera is one of the most important technology of self-driving cars.It is also a significant application of artificial intelligence in computer vision.Traditional traffic light detection algorithms used to use color segmentation to detect traffic light bulb or use sliding windows to find traffic light.But these two methods are either too inaccurate or too time consuming.The state-of-the-art object detection algorithms based on deep convolutional neural network are not suitable for small objects like traffic lights.Therefore,in order to solve the time consuming problem and improve accuracy,this thesis proposes an accurate and efficient traffic light detection algorithm based on color segmentation and deep convolutional neural network.The main work of this thesis includes the following three parts:1.An algorithm of extracting common feature areas of different types of traffic lights is proposed.The traffic light bulb is regarded as the common feature area of different types of traffic lights.By using color segmentation method,image eroding,image dilating and connection domain extraction,the traffic light bulbs can be found.For most of the traffic light bulbs are in the top or bottom of the traffic lights,so the possible locations of the traffic lights can be quickly located.2.An algorithm of generating traffic light candidate areas is proposed.Generally,there is a mapping relationship between the traffic light bulb and the traffic light.This algorithm establishes a model which can generate the traffic light candidate areas from the traffic light bulb and optimizes the model parameters using mathematical statistics and genetic algorithm.In order to make the generated traffic light candidate areas cover the traffic light locations sufficiently,one traffic light bulb need to generate a number of traffic light candidate areas.3.An algorithm of traffic light classification is proposed.The amount of the traffic light candidate areas is very large and not all of them contains traffic lights,so it is necessary to design a classification algorithm to classify the candidate areas.The classification algorithm is based on Fast RCNN.By improving the network architecture of Fast RCNN,the classification algorithm is more suitable for traffic light detection.